| from torch import nn |
| import torch |
| from modules import devices, shared |
| import ldm.models.diffusion.ddpm |
|
|
|
|
| class VectorAdjustPrior(nn.Module): |
| def __init__(self, hidden_size, inter_dim=64): |
| super().__init__() |
| self.vector_proj = nn.Linear(hidden_size * 2, inter_dim, bias=True) |
| self.out_proj = nn.Linear(hidden_size + inter_dim, hidden_size, bias=True) |
|
|
| def forward(self, z): |
| b, s = z.shape[0:2] |
| x1 = torch.mean(z, dim=1).repeat(s, 1) |
| x2 = z.reshape(b * s, -1) |
| x = torch.cat((x1, x2), dim=1) |
| x = self.vector_proj(x) |
| x = torch.cat((x2, x), dim=1) |
| x = self.out_proj(x) |
| x = x.reshape(b, s, -1) |
| return x |
|
|
| @classmethod |
| def load_model(cls, model_path, hidden_size=768, inter_dim=64): |
| model = cls(hidden_size=hidden_size, inter_dim=inter_dim) |
| model.load_state_dict(torch.load(model_path)["state_dict"]) |
|
|
| return model |
|
|
|
|
| vap = VectorAdjustPrior.load_model('v2.pt').to(devices.device) |
|
|
|
|
| def get_learned_conditioning_with_prior(self, c): |
| cond = ldm.models.diffusion.ddpm.LatentDiffusion.get_learned_conditioning_original(self, c) |
|
|
| if shared.opts.use_prior: |
| cond = vap(cond) |
|
|
| return cond |
|
|
|
|
| if not hasattr(ldm.models.diffusion.ddpm.LatentDiffusion, 'get_learned_conditioning_original'): |
| ldm.models.diffusion.ddpm.LatentDiffusion.get_learned_conditioning_original = ldm.models.diffusion.ddpm.LatentDiffusion.get_learned_conditioning |
| ldm.models.diffusion.ddpm.LatentDiffusion.get_learned_conditioning = get_learned_conditioning_with_prior |
|
|
| shared.options_templates.update(shared.options_section(('nai', "NAI"), { |
| "use_prior": shared.OptionInfo(False, "use v2.pt"), |
| })) |
